├── .gitignore ├── LICENSE ├── README.md ├── data_generator.py ├── data_preprocess.py ├── emph.py ├── emph_test.py ├── model.py ├── requirements.txt └── vbnorm.py /.gitignore: -------------------------------------------------------------------------------- 1 | # python 2 | __pycache__/ 3 | # python virtual environment 4 | venv/ 5 | 6 | # project specifics 7 | log/ 8 | gen_data*/ 9 | models/ 10 | 11 | # swap files for vi editor 12 | *.swo 13 | *.swp 14 | 15 | # IntelliJ IDEA users 16 | .idea/ 17 | 18 | # OSX 19 | .DS_Store 20 | 21 | # data paths 22 | segan_data_in/ 23 | segan_data_out/ 24 | 25 | # c3dl-specific (navercorp) 26 | run.sh 27 | segan.json 28 | train.sh 29 | user_dev_workspace/ 30 | run_attach.sh 31 | 32 | # tags file 33 | tags 34 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Pytorch Implementation of SEGAN (Speech Enhancement GAN) 2 | Implementation of [SEGAN](https://arxiv.org/abs/1703.09452) by Pascual et al. in 2017, using pytorch. 3 | Original Tensorflow version can be found [here](https://github.com/santi-pdp/segan). 4 | 5 | ## Prerequisites 6 | 7 | - python v3.5.2 or higher 8 | - pytorch v0.4.0 9 | - CUDA preferred 10 | - noisy speech dataset downloaded from [here](https://datashare.is.ed.ac.uk/handle/10283/2791) 11 | - libraries specified in `requirements.txt` 12 | 13 | ### Installing Required Libraries 14 | 15 | `pip install -r requirements.txt` 16 | 17 | ## Data Preprocessing 18 | 19 | Use `data_preprocess.py` file to preprocess downloaded data. 20 | Adjust the file paths at the beginning of the file to properly locate the data files, output folder, etc. 21 | Uncomment functions in `__main__` to perform desired preprocessing stage. 22 | 23 | Data preprocessing consists of three main stages: 24 | 1. Downsampling - downsample original audio files (48k) to sampling rate of 16000. 25 | 2. Serialization - Splitting the audio files into 2^14-sample (about 1 second) snippets. 26 | 3. Verification - whether it contains proper number of samples. 27 | 28 | Note that the second stage takes a fairly long time - more than an hour. 29 | 30 | ## Training 31 | 32 | ``` 33 | python model.py 34 | ``` 35 | 36 | Again, fix and adjust datapaths in `model.py` according to your needs. 37 | Especially, provide accurate path to where serialized data are stored. 38 | 39 | ## Using Tensorboard 40 | 41 | In order to use tensorboard, you need to first install tensorboard: 42 | 43 | ``` 44 | pip install tensorboard 45 | ``` 46 | 47 | Then run tensorboard by specifing the log directory. 48 | 49 | ``` 50 | tensorboard --logdir=segan_data_out/tblogs 51 | ``` 52 | -------------------------------------------------------------------------------- /data_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.utils import data 3 | import numpy as np 4 | import os 5 | 6 | 7 | class AudioSampleGenerator(data.Dataset): 8 | """ 9 | Audio sample reader. 10 | Used alongside with DataLoader class to generate batches. 11 | see: http://pytorch.org/docs/master/data.html#torch.utils.data.Dataset 12 | """ 13 | SAMPLE_LENGTH = 16384 14 | 15 | def __init__(self, data_folder_path: str): 16 | if not os.path.exists(data_folder_path): 17 | raise FileNotFoundError 18 | 19 | # store full paths - not the actual files. 20 | # all files cannot be loaded up to memory due to its large size. 21 | # insted, we read from files upon fetching batches (see __getitem__() implementation) 22 | self.filepaths = [os.path.join(data_folder_path, filename) 23 | for filename in os.listdir(data_folder_path)] 24 | self.num_data = len(self.filepaths) 25 | 26 | def reference_batch(self, batch_size: int): 27 | """ 28 | Randomly selects a reference batch from dataset. 29 | Reference batch is used for calculating statistics for virtual batch normalization operation. 30 | 31 | Args: 32 | batch_size(int): batch size 33 | 34 | Returns: 35 | ref_batch: reference batch 36 | """ 37 | ref_filenames = np.random.choice(self.filepaths, batch_size) 38 | ref_batch = torch.from_numpy(np.stack([np.load(f) for f in ref_filenames])) 39 | return ref_batch 40 | 41 | def fixed_test_audio(self, num_test_audio: int): 42 | """ 43 | Randomly chosen batch for testing generated results. 44 | 45 | Args: 46 | num_test_audio(int): number of test audio. 47 | Must be same as batch size of training, 48 | otherwise it cannot go through the forward step of generator. 49 | """ 50 | test_filenames = np.random.choice(self.filepaths, num_test_audio) 51 | # stack the data for all test audios 52 | test_audios = np.stack([np.load(f) for f in test_filenames]) 53 | test_clean_set = test_audios[:, 0].reshape((num_test_audio, 1, self.SAMPLE_LENGTH)) 54 | test_noisy_set = test_audios[:, 1].reshape((num_test_audio, 1, self.SAMPLE_LENGTH)) 55 | # file names of test samples 56 | test_basenames = [os.path.basename(fpath) for fpath in test_filenames] 57 | return test_basenames, test_clean_set, test_noisy_set 58 | 59 | def __getitem__(self, idx): 60 | # get item for specified index 61 | pair = np.load(self.filepaths[idx]) 62 | return pair 63 | 64 | def __len__(self): 65 | return self.num_data 66 | 67 | -------------------------------------------------------------------------------- /data_preprocess.py: -------------------------------------------------------------------------------- 1 | import os 2 | import subprocess 3 | import librosa 4 | import numpy as np 5 | import time 6 | 7 | 8 | """ 9 | Audio data preprocessing for SEGAN training. 10 | 11 | It provides: 12 | 1. 16k downsampling (sox required) 13 | 2. slicing and serializing 14 | 3. verifying serialized data 15 | """ 16 | 17 | 18 | # specify the paths - modify the paths at your will 19 | DATA_ROOT_DIR = '../data/segan' # the base folder for dataset 20 | CLEAN_TRAIN_DIR = 'clean_trainset_56spk_wav' # where original clean train data exist 21 | NOISY_TRAIN_DIR = 'noisy_trainset_56spk_wav' # where original noisy train data exist 22 | DST_CLEAN_TRAIN_DIR = 'clean_trainset_wav_16k' # clean preprocessed data folder 23 | DST_NOISY_TRAIN_DIR = 'noisy_trainset_wav_16k' # noisy preprocessed data folder 24 | SER_DATA_DIR = 'ser_data' # serialized data folder 25 | SER_DST_PATH = os.path.join(DATA_ROOT_DIR, SER_DATA_DIR) 26 | 27 | 28 | def verify_data(): 29 | """ 30 | Verifies the length of each data after preprocessing. 31 | """ 32 | for dirname, dirs, files in os.walk(SER_DST_PATH): 33 | for filename in files: 34 | data_pair = np.load(os.path.join(dirname, filename)) 35 | if data_pair.shape[1] != 16384: 36 | print('Snippet length not 16384 : {} instead'.format(data_pair.shape[1])) 37 | break 38 | 39 | 40 | def downsample_16k(): 41 | """ 42 | Convert all audio files to have sampling rate 16k. 43 | """ 44 | # clean training sets 45 | dst_clean_dir = os.path.join(DATA_ROOT_DIR, DST_CLEAN_TRAIN_DIR) 46 | if not os.path.exists(dst_clean_dir): 47 | os.makedirs(dst_clean_dir) 48 | 49 | for dirname, dirs, files in os.walk(os.path.join(DATA_ROOT_DIR, CLEAN_TRAIN_DIR)): 50 | for filename in files: 51 | input_filepath = os.path.abspath(os.path.join(dirname, filename)) 52 | out_filepath = os.path.join(dst_clean_dir, filename) 53 | # use sox to down-sample to 16k 54 | print('Downsampling : {}'.format(input_filepath)) 55 | subprocess.run( 56 | 'sox {} -r 16k {}' 57 | .format(input_filepath, out_filepath), 58 | shell=True, check=True) 59 | 60 | # noisy training sets 61 | dst_noisy_dir = os.path.join(DATA_ROOT_DIR, DST_NOISY_TRAIN_DIR) 62 | if not os.path.exists(dst_noisy_dir): 63 | os.makedirs(dst_noisy_dir) 64 | 65 | for dirname, dirs, files in os.walk(os.path.join(DATA_ROOT_DIR, NOISY_TRAIN_DIR)): 66 | for filename in files: 67 | input_filepath = os.path.abspath(os.path.join(dirname, filename)) 68 | out_filepath = os.path.join(dst_noisy_dir, filename) 69 | # use sox to down-sample to 16k 70 | print('Processing : {}'.format(input_filepath)) 71 | subprocess.run( 72 | 'sox {} -r 16k {}' 73 | .format(input_filepath, out_filepath), 74 | shell=True, check=True) 75 | 76 | 77 | def slice_signal(filepath, window_size, stride, sample_rate): 78 | """ 79 | Helper function for slicing the audio file 80 | by window size with [stride] percent overlap (default 50%). 81 | """ 82 | wav, sr = librosa.load(filepath, sr=sample_rate) 83 | n_samples = wav.shape[0] # contains simple amplitudes 84 | hop = int(window_size * stride) 85 | slices = [] 86 | for end_idx in range(window_size, len(wav), hop): 87 | start_idx = end_idx - window_size 88 | slice_sig = wav[start_idx:end_idx] 89 | slices.append(slice_sig) 90 | return slices 91 | 92 | 93 | def process_and_serialize(): 94 | """ 95 | Serialize the sliced signals and save on separate folder. 96 | """ 97 | start_time = time.time() # measure the time 98 | window_size = 2 ** 14 # about 1 second of samples 99 | sample_rate = 16000 100 | stride = 0.5 101 | 102 | if not os.path.exists(SER_DST_PATH): 103 | print('Creating new destination folder for new data') 104 | os.makedirs(SER_DST_PATH) 105 | 106 | # the path for source data (16k downsampled) 107 | clean_data_path = os.path.join(DATA_ROOT_DIR, DST_CLEAN_TRAIN_DIR) 108 | noisy_data_path = os.path.join(DATA_ROOT_DIR, DST_NOISY_TRAIN_DIR) 109 | 110 | # walk through the path, slice the audio file, and save the serialized result 111 | for dirname, dirs, files in os.walk(clean_data_path): 112 | if len(files) == 0: 113 | continue 114 | for filename in files: 115 | print('Splitting : {}'.format(filename)) 116 | clean_filepath = os.path.join(clean_data_path, filename) 117 | noisy_filepath = os.path.join(noisy_data_path, filename) 118 | 119 | # slice both clean signal and noisy signal 120 | clean_sliced = slice_signal(clean_filepath, window_size, stride, sample_rate) 121 | noisy_sliced = slice_signal(noisy_filepath, window_size, stride, sample_rate) 122 | 123 | # serialize - file format goes [original_file]_[slice_number].npy 124 | # ex) p293_154.wav_5.npy denotes 5th slice of p293_154.wav file 125 | for idx, slice_tuple in enumerate(zip(clean_sliced, noisy_sliced)): 126 | pair = np.array([slice_tuple[0], slice_tuple[1]]) 127 | np.save(os.path.join(SER_DST_PATH, '{}_{}'.format(filename, idx)), arr=pair) 128 | 129 | # measure the time it took to process 130 | end_time = time.time() 131 | print('Total elapsed time for preprocessing : {}'.format(end_time - start_time)) 132 | 133 | 134 | if __name__ == '__main__': 135 | """ 136 | Uncomment each function call that suits your needs. 137 | """ 138 | # downsample_16k() 139 | # process_and_serialize() # WARNING - takes very long time 140 | # verify_data() 141 | -------------------------------------------------------------------------------- /emph.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from scipy import signal 3 | 4 | 5 | def pre_emphasis(signal_batch, emph_coeff=0.95) -> np.array: 6 | """ 7 | Pre-emphasis of higher frequencies given a batch of signal. 8 | 9 | Args: 10 | signal_batch(np.array): batch of signals, represented as numpy arrays 11 | emph_coeff(float): emphasis coefficient 12 | 13 | Returns: 14 | result: pre-emphasized signal batch 15 | """ 16 | return signal.lfilter([1, -emph_coeff], [1], signal_batch) 17 | 18 | 19 | def de_emphasis(signal_batch, emph_coeff=0.95) -> np.array: 20 | """ 21 | De-emphasis operation given a batch of signal. 22 | Reverts the pre-emphasized signal. 23 | 24 | Args: 25 | signal_batch(np.array): batch of signals, represented as numpy arrays 26 | emph_coeff(float): emphasis coefficient 27 | 28 | Returns: 29 | result: de-emphasized signal batch 30 | """ 31 | return signal.lfilter([1], [1, -emph_coeff], signal_batch) 32 | -------------------------------------------------------------------------------- /emph_test.py: -------------------------------------------------------------------------------- 1 | import unittest 2 | import numpy as np 3 | import emph 4 | 5 | 6 | class TestEmphasis(unittest.TestCase): 7 | def test_pre_emphasis(self): 8 | """ 9 | Tests equality after de-emphasizing pre-emphasized signal. 10 | """ 11 | rand_signal_batch = np.random.randint(low=1, high=10, size=(10, 1, 400)) 12 | reconst_batch = emph.de_emphasis(emph.pre_emphasis(rand_signal_batch)) 13 | 14 | # after de-emphasis, the signal must have been restored 15 | self.assertEqual(rand_signal_batch.shape, reconst_batch.shape) 16 | self.assertTrue(np.allclose(rand_signal_batch, reconst_batch)) 17 | 18 | 19 | if __name__ == '__main__': 20 | unittest.main() 21 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | """ 2 | Here we define the discriminator and generator for SEGAN. 3 | After definition of each modules, run the training. 4 | """ 5 | 6 | import time 7 | import os 8 | import torch 9 | from torch import nn 10 | from torch.utils.data import DataLoader 11 | from torch import optim 12 | import numpy as np 13 | from scipy.io import wavfile 14 | from data_generator import AudioSampleGenerator 15 | from vbnorm import VirtualBatchNorm1d 16 | from tensorboardX import SummaryWriter 17 | import emph 18 | 19 | # device we're using 20 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 21 | 22 | # define folders for output data 23 | in_path = 'segan_data_in' 24 | out_path_root = 'segan_data_out' 25 | ser_data_fdr = 'ser_data' # serialized data 26 | gen_data_fdr = 'gen_data' # folder for saving generated data 27 | checkpoint_fdr = 'checkpoint' # folder for saving models, optimizer states, etc. 28 | tblog_fdr = 'logs' # summary data for tensorboard 29 | # time info is used to distinguish dfferent training sessions 30 | run_time = time.strftime('%Y%m%d_%H%M', time.gmtime()) # 20180625_1742 31 | # output path - all outputs (generated data, logs, model checkpoints) will be stored here 32 | # the directory structure is as: "[curr_dir]/segan_data_out/[run_time]/" 33 | out_path = os.path.join(os.getcwd(), out_path_root, run_time) 34 | tblog_path = os.path.join(os.getcwd(), tblog_fdr, run_time) # summary data for tensorboard 35 | 36 | 37 | # create folder for generated data 38 | gen_data_path = os.path.join(out_path, gen_data_fdr) 39 | if not os.path.exists(gen_data_path): 40 | os.makedirs(gen_data_path) 41 | 42 | # create folder for model checkpoints 43 | checkpoint_path = os.path.join(out_path, checkpoint_fdr) 44 | if not os.path.exists(checkpoint_path): 45 | os.makedirs(checkpoint_path) 46 | 47 | 48 | class Discriminator(nn.Module): 49 | """D""" 50 | def __init__(self, dropout_drop=0.5): 51 | super().__init__() 52 | # Define convolution operations. 53 | # (#input channel, #output channel, kernel_size, stride, padding) 54 | # in : 16384 x 2 55 | negative_slope = 0.03 56 | self.conv1 = nn.Conv1d(in_channels=2, out_channels=32, kernel_size=31, stride=2, padding=15) # out : 8192 x 32 57 | self.vbn1 = VirtualBatchNorm1d(32) 58 | self.lrelu1 = nn.LeakyReLU(negative_slope) 59 | self.conv2 = nn.Conv1d(32, 64, 31, 2, 15) # 4096 x 64 60 | self.vbn2 = VirtualBatchNorm1d(64) 61 | self.lrelu2 = nn.LeakyReLU(negative_slope) 62 | self.conv3 = nn.Conv1d(64, 64, 31, 2, 15) # 2048 x 64 63 | self.dropout1 = nn.Dropout(dropout_drop) 64 | self.vbn3 = VirtualBatchNorm1d(64) 65 | self.lrelu3 = nn.LeakyReLU(negative_slope) 66 | self.conv4 = nn.Conv1d(64, 128, 31, 2, 15) # 1024 x 128 67 | self.vbn4 = VirtualBatchNorm1d(128) 68 | self.lrelu4 = nn.LeakyReLU(negative_slope) 69 | self.conv5 = nn.Conv1d(128, 128, 31, 2, 15) # 512 x 128 70 | self.vbn5 = VirtualBatchNorm1d(128) 71 | self.lrelu5 = nn.LeakyReLU(negative_slope) 72 | self.conv6 = nn.Conv1d(128, 256, 31, 2, 15) # 256 x 256 73 | self.dropout2 = nn.Dropout(dropout_drop) 74 | self.vbn6 = VirtualBatchNorm1d(256) 75 | self.lrelu6 = nn.LeakyReLU(negative_slope) 76 | self.conv7 = nn.Conv1d(256, 256, 31, 2, 15) # 128 x 256 77 | self.vbn7 = VirtualBatchNorm1d(256) 78 | self.lrelu7 = nn.LeakyReLU(negative_slope) 79 | self.conv8 = nn.Conv1d(256, 512, 31, 2, 15) # 64 x 512 80 | self.vbn8 = VirtualBatchNorm1d(512) 81 | self.lrelu8 = nn.LeakyReLU(negative_slope) 82 | self.conv9 = nn.Conv1d(512, 512, 31, 2, 15) # 32 x 512 83 | self.dropout3 = nn.Dropout(dropout_drop) 84 | self.vbn9 = VirtualBatchNorm1d(512) 85 | self.lrelu9 = nn.LeakyReLU(negative_slope) 86 | self.conv10 = nn.Conv1d(512, 1024, 31, 2, 15) # 16 x 1024 87 | self.vbn10 = VirtualBatchNorm1d(1024) 88 | self.lrelu10 = nn.LeakyReLU(negative_slope) 89 | self.conv11 = nn.Conv1d(1024, 2048, 31, 2, 15) # 8 x 1024 90 | self.vbn11 = VirtualBatchNorm1d(2048) 91 | self.lrelu11 = nn.LeakyReLU(negative_slope) 92 | # 1x1 size kernel for dimension and parameter reduction 93 | self.conv_final = nn.Conv1d(2048, 1, kernel_size=1, stride=1) # 8 x 1 94 | self.lrelu_final = nn.LeakyReLU(negative_slope) 95 | self.fully_connected = nn.Linear(in_features=8, out_features=1) # 1 96 | self.sigmoid = nn.Sigmoid() 97 | 98 | # initialize weights 99 | self.init_weights() 100 | 101 | def init_weights(self): 102 | """ 103 | Initialize weights for convolution layers using Xavier initialization. 104 | """ 105 | for m in self.modules(): 106 | if isinstance(m, nn.Conv1d): 107 | nn.init.xavier_normal_(m.weight.data) 108 | 109 | def forward(self, x, ref_x): 110 | """ 111 | Forward pass of discriminator. 112 | 113 | Args: 114 | x: batch 115 | ref_x: reference batch for virtual batch norm 116 | """ 117 | # reference pass 118 | ref_x = self.conv1(ref_x) 119 | ref_x, mean1, meansq1 = self.vbn1(ref_x, None, None) 120 | ref_x = self.lrelu1(ref_x) 121 | ref_x = self.conv2(ref_x) 122 | ref_x, mean2, meansq2 = self.vbn2(ref_x, None, None) 123 | ref_x = self.lrelu2(ref_x) 124 | ref_x = self.conv3(ref_x) 125 | ref_x = self.dropout1(ref_x) 126 | ref_x, mean3, meansq3 = self.vbn3(ref_x, None, None) 127 | ref_x = self.lrelu3(ref_x) 128 | ref_x = self.conv4(ref_x) 129 | ref_x, mean4, meansq4 = self.vbn4(ref_x, None, None) 130 | ref_x = self.lrelu4(ref_x) 131 | ref_x = self.conv5(ref_x) 132 | ref_x, mean5, meansq5 = self.vbn5(ref_x, None, None) 133 | ref_x = self.lrelu5(ref_x) 134 | ref_x = self.conv6(ref_x) 135 | ref_x = self.dropout2(ref_x) 136 | ref_x, mean6, meansq6 = self.vbn6(ref_x, None, None) 137 | ref_x = self.lrelu6(ref_x) 138 | ref_x = self.conv7(ref_x) 139 | ref_x, mean7, meansq7 = self.vbn7(ref_x, None, None) 140 | ref_x = self.lrelu7(ref_x) 141 | ref_x = self.conv8(ref_x) 142 | ref_x, mean8, meansq8 = self.vbn8(ref_x, None, None) 143 | ref_x = self.lrelu8(ref_x) 144 | ref_x = self.conv9(ref_x) 145 | ref_x = self.dropout3(ref_x) 146 | ref_x, mean9, meansq9 = self.vbn9(ref_x, None, None) 147 | ref_x = self.lrelu9(ref_x) 148 | ref_x = self.conv10(ref_x) 149 | ref_x, mean10, meansq10 = self.vbn10(ref_x, None, None) 150 | ref_x = self.lrelu10(ref_x) 151 | ref_x = self.conv11(ref_x) 152 | ref_x, mean11, meansq11 = self.vbn11(ref_x, None, None) 153 | # further pass no longer needed 154 | 155 | # train pass 156 | x = self.conv1(x) 157 | x, _, _ = self.vbn1(x, mean1, meansq1) 158 | x = self.lrelu1(x) 159 | x = self.conv2(x) 160 | x, _, _ = self.vbn2(x, mean2, meansq2) 161 | x = self.lrelu2(x) 162 | x = self.conv3(x) 163 | x = self.dropout1(x) 164 | x, _, _ = self.vbn3(x, mean3, meansq3) 165 | x = self.lrelu3(x) 166 | x = self.conv4(x) 167 | x, _, _ = self.vbn4(x, mean4, meansq4) 168 | x = self.lrelu4(x) 169 | x = self.conv5(x) 170 | x, _, _ = self.vbn5(x, mean5, meansq5) 171 | x = self.lrelu5(x) 172 | x = self.conv6(x) 173 | x = self.dropout2(x) 174 | x, _, _ = self.vbn6(x, mean6, meansq6) 175 | x = self.lrelu6(x) 176 | x = self.conv7(x) 177 | x, _, _ = self.vbn7(x, mean7, meansq7) 178 | x = self.lrelu7(x) 179 | x = self.conv8(x) 180 | x, _, _ = self.vbn8(x, mean8, meansq8) 181 | x = self.lrelu8(x) 182 | x = self.conv9(x) 183 | x = self.dropout3(x) 184 | x, _, _ = self.vbn9(x, mean9, meansq9) 185 | x = self.lrelu9(x) 186 | x = self.conv10(x) 187 | x, _, _ = self.vbn10(x, mean10, meansq10) 188 | x = self.lrelu10(x) 189 | x = self.conv11(x) 190 | x, _, _ = self.vbn11(x, mean11, meansq11) 191 | x = self.lrelu11(x) 192 | x = self.conv_final(x) 193 | x = self.lrelu_final(x) 194 | # reduce down to a scalar value 195 | x = torch.squeeze(x) 196 | x = self.fully_connected(x) 197 | # return self.sigmoid(x) 198 | return x 199 | 200 | 201 | class Generator(nn.Module): 202 | """G""" 203 | def __init__(self): 204 | super().__init__() 205 | # size notations = [batch_size x feature_maps x width] (height omitted - 1D convolutions) 206 | # encoder gets a noisy signal as input 207 | self.enc1 = nn.Conv1d(in_channels=1, out_channels=16, kernel_size=32, stride=2, padding=15) # out : [B x 16 x 8192] 208 | self.enc1_nl = nn.PReLU() # non-linear transformation after encoder layer 1 209 | self.enc2 = nn.Conv1d(16, 32, 32, 2, 15) # [B x 32 x 4096] 210 | self.enc2_nl = nn.PReLU() 211 | self.enc3 = nn.Conv1d(32, 32, 32, 2, 15) # [B x 32 x 2048] 212 | self.enc3_nl = nn.PReLU() 213 | self.enc4 = nn.Conv1d(32, 64, 32, 2, 15) # [B x 64 x 1024] 214 | self.enc4_nl = nn.PReLU() 215 | self.enc5 = nn.Conv1d(64, 64, 32, 2, 15) # [B x 64 x 512] 216 | self.enc5_nl = nn.PReLU() 217 | self.enc6 = nn.Conv1d(64, 128, 32, 2, 15) # [B x 128 x 256] 218 | self.enc6_nl = nn.PReLU() 219 | self.enc7 = nn.Conv1d(128, 128, 32, 2, 15) # [B x 128 x 128] 220 | self.enc7_nl = nn.PReLU() 221 | self.enc8 = nn.Conv1d(128, 256, 32, 2, 15) # [B x 256 x 64] 222 | self.enc8_nl = nn.PReLU() 223 | self.enc9 = nn.Conv1d(256, 256, 32, 2, 15) # [B x 256 x 32] 224 | self.enc9_nl = nn.PReLU() 225 | self.enc10 = nn.Conv1d(256, 512, 32, 2, 15) # [B x 512 x 16] 226 | self.enc10_nl = nn.PReLU() 227 | self.enc11 = nn.Conv1d(512, 1024, 32, 2, 15) # output : [B x 1024 x 8] 228 | self.enc11_nl = nn.PReLU() 229 | 230 | # decoder generates an enhanced signal 231 | # each decoder output are concatenated with homolgous encoder output, 232 | # so the feature map sizes are doubled 233 | self.dec10 = nn.ConvTranspose1d(in_channels=2048, out_channels=512, kernel_size=32, stride=2, padding=15) 234 | self.dec10_nl = nn.PReLU() # out : [B x 512 x 16] -> (concat) [B x 1024 x 16] 235 | self.dec9 = nn.ConvTranspose1d(1024, 256, 32, 2, 15) # [B x 256 x 32] 236 | self.dec9_nl = nn.PReLU() 237 | self.dec8 = nn.ConvTranspose1d(512, 256, 32, 2, 15) # [B x 256 x 64] 238 | self.dec8_nl = nn.PReLU() 239 | self.dec7 = nn.ConvTranspose1d(512, 128, 32, 2, 15) # [B x 128 x 128] 240 | self.dec7_nl = nn.PReLU() 241 | self.dec6 = nn.ConvTranspose1d(256, 128, 32, 2, 15) # [B x 128 x 256] 242 | self.dec6_nl = nn.PReLU() 243 | self.dec5 = nn.ConvTranspose1d(256, 64, 32, 2, 15) # [B x 64 x 512] 244 | self.dec5_nl = nn.PReLU() 245 | self.dec4 = nn.ConvTranspose1d(128, 64, 32, 2, 15) # [B x 64 x 1024] 246 | self.dec4_nl = nn.PReLU() 247 | self.dec3 = nn.ConvTranspose1d(128, 32, 32, 2, 15) # [B x 32 x 2048] 248 | self.dec3_nl = nn.PReLU() 249 | self.dec2 = nn.ConvTranspose1d(64, 32, 32, 2, 15) # [B x 32 x 4096] 250 | self.dec2_nl = nn.PReLU() 251 | self.dec1 = nn.ConvTranspose1d(64, 16, 32, 2, 15) # [B x 16 x 8192] 252 | self.dec1_nl = nn.PReLU() 253 | self.dec_final = nn.ConvTranspose1d(32, 1, 32, 2, 15) # [B x 1 x 16384] 254 | self.dec_tanh = nn.Tanh() 255 | 256 | # initialize weights 257 | self.init_weights() 258 | 259 | def init_weights(self): 260 | """ 261 | Initialize weights for convolution layers using Xavier initialization. 262 | """ 263 | for m in self.modules(): 264 | if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d): 265 | nn.init.xavier_normal_(m.weight.data) 266 | 267 | def forward(self, x, z): 268 | """ 269 | Forward pass of generator. 270 | 271 | Args: 272 | x: input batch (signal) 273 | z: latent vector 274 | """ 275 | ### encoding step 276 | e1 = self.enc1(x) 277 | e2 = self.enc2(self.enc1_nl(e1)) 278 | e3 = self.enc3(self.enc2_nl(e2)) 279 | e4 = self.enc4(self.enc3_nl(e3)) 280 | e5 = self.enc5(self.enc4_nl(e4)) 281 | e6 = self.enc6(self.enc5_nl(e5)) 282 | e7 = self.enc7(self.enc6_nl(e6)) 283 | e8 = self.enc8(self.enc7_nl(e7)) 284 | e9 = self.enc9(self.enc8_nl(e8)) 285 | e10 = self.enc10(self.enc9_nl(e9)) 286 | e11 = self.enc11(self.enc10_nl(e10)) 287 | # c = compressed feature, the 'thought vector' 288 | c = self.enc11_nl(e11) 289 | 290 | # concatenate the thought vector with latent variable 291 | encoded = torch.cat((c, z), dim=1) 292 | 293 | ### decoding step 294 | d10 = self.dec10(encoded) 295 | # dx_c : concatenated with skip-connected layer's output & passed nonlinear layer 296 | d10_c = self.dec10_nl(torch.cat((d10, e10), dim=1)) 297 | d9 = self.dec9(d10_c) 298 | d9_c = self.dec9_nl(torch.cat((d9, e9), dim=1)) 299 | d8 = self.dec8(d9_c) 300 | d8_c = self.dec8_nl(torch.cat((d8, e8), dim=1)) 301 | d7 = self.dec7(d8_c) 302 | d7_c = self.dec7_nl(torch.cat((d7, e7), dim=1)) 303 | d6 = self.dec6(d7_c) 304 | d6_c = self.dec6_nl(torch.cat((d6, e6), dim=1)) 305 | d5 = self.dec5(d6_c) 306 | d5_c = self.dec5_nl(torch.cat((d5, e5), dim=1)) 307 | d4 = self.dec4(d5_c) 308 | d4_c = self.dec4_nl(torch.cat((d4, e4), dim=1)) 309 | d3 = self.dec3(d4_c) 310 | d3_c = self.dec3_nl(torch.cat((d3, e3), dim=1)) 311 | d2 = self.dec2(d3_c) 312 | d2_c = self.dec2_nl(torch.cat((d2, e2), dim=1)) 313 | d1 = self.dec1(d2_c) 314 | d1_c = self.dec1_nl(torch.cat((d1, e1), dim=1)) 315 | out = self.dec_tanh(self.dec_final(d1_c)) 316 | return out 317 | 318 | 319 | def split_pair_to_vars(sample_batch_pair): 320 | """ 321 | Splits the generated batch data and creates combination of pairs. 322 | Input argument sample_batch_pair consists of a batch_size number of 323 | [clean_signal, noisy_signal] pairs. 324 | 325 | This function creates three pytorch Variables - a clean_signal, noisy_signal pair, 326 | clean signal only, and noisy signal only. 327 | It goes through preemphasis preprocessing before converted into variable. 328 | 329 | Args: 330 | sample_batch_pair(torch.Tensor): batch of [clean_signal, noisy_signal] pairs 331 | Returns: 332 | batch_pairs_var(Variable): batch of pairs containing clean signal and noisy signal 333 | clean_batch_var(Variable): clean signal batch 334 | noisy_batch_var(Varialbe): noisy signal batch 335 | """ 336 | # pre-emphasis 337 | sample_batch_pair = emph.pre_emphasis(sample_batch_pair.numpy(), emph_coeff=0.95) 338 | 339 | batch_pairs_var = torch.from_numpy(sample_batch_pair).type(torch.FloatTensor).to(device) # [40 x 2 x 16384] 340 | clean_batch = np.stack([pair[0].reshape(1, -1) for pair in sample_batch_pair]) 341 | clean_batch_var = torch.from_numpy(clean_batch).type(torch.FloatTensor).to(device) 342 | noisy_batch = np.stack([pair[1].reshape(1, -1) for pair in sample_batch_pair]) 343 | noisy_batch_var = torch.from_numpy(noisy_batch).type(torch.FloatTensor).to(device) 344 | return batch_pairs_var, clean_batch_var, noisy_batch_var 345 | 346 | 347 | def sample_latent(): 348 | """ 349 | Sample a latent vector - normal distribution 350 | 351 | Returns: 352 | z(torch.Tensor): random latent vector 353 | """ 354 | return torch.randn((batch_size, 1024, 8)).to(device) 355 | 356 | 357 | # SOME TRAINING PARAMETERS # 358 | batch_size = 128 359 | d_learning_rate = 0.0001 360 | g_learning_rate = 0.0001 361 | g_lambda = 100 # regularizer for generator 362 | use_devices = [0, 1, 2, 3] 363 | sample_rate = 16000 364 | num_gen_examples = 10 # number of generated audio examples displayed per epoch 365 | num_epochs = 86 366 | 367 | # create D and G instances 368 | discriminator = torch.nn.DataParallel(Discriminator().to(device), device_ids=use_devices) # use GPU 369 | print(discriminator) 370 | print('Discriminator created') 371 | 372 | generator = torch.nn.DataParallel(Generator().to(device), device_ids=use_devices) 373 | print(generator) 374 | print('Generator created') 375 | 376 | # This is how you define a data loader 377 | sample_generator = AudioSampleGenerator(os.path.join(in_path, ser_data_fdr)) 378 | random_data_loader = DataLoader( 379 | dataset=sample_generator, 380 | batch_size=batch_size, # specified batch size here 381 | shuffle=True, 382 | num_workers=4, 383 | drop_last=True, # drop the last batch that cannot be divided by batch_size 384 | pin_memory=True) 385 | print('DataLoader created') 386 | 387 | # generate reference batch 388 | ref_batch_pairs = sample_generator.reference_batch(batch_size) 389 | ref_batch_var, ref_clean_var, ref_noisy_var = split_pair_to_vars(ref_batch_pairs) 390 | 391 | # optimizers 392 | g_optimizer = optim.Adam(generator.parameters(), lr=g_learning_rate, betas=(0.5, 0.999)) 393 | d_optimizer = optim.Adam(discriminator.parameters(), lr=d_learning_rate, betas=(0.5, 0.999)) 394 | 395 | # create tensorboard writer 396 | # The logs will be stored NOT under the run_time, but under segan_data_out/'tblog_fdr'. 397 | # This way, tensorboard can show graphs for each experiment in one board 398 | tbwriter = SummaryWriter(log_dir=tblog_path) 399 | print('TensorboardX summary writer created') 400 | 401 | # test samples for generation 402 | test_noise_filenames, fixed_test_clean, fixed_test_noise = \ 403 | sample_generator.fixed_test_audio(num_gen_examples) 404 | fixed_test_clean = torch.from_numpy(fixed_test_clean) 405 | fixed_test_noise = torch.from_numpy(fixed_test_noise) 406 | print('Test samples loaded') 407 | 408 | # record the fixed examples 409 | for idx, fname in enumerate(test_noise_filenames): 410 | tbwriter.add_audio( 411 | 'test_audio_clean/{}'.format(fname), 412 | fixed_test_clean.numpy()[idx].T, 413 | sample_rate=sample_rate) 414 | tbwriter.add_audio( 415 | 'test_audio_noise/{}'.format(fname), 416 | fixed_test_noise.numpy()[idx].T, 417 | sample_rate=sample_rate) 418 | 419 | 420 | ### Train! ### 421 | print('Starting Training...') 422 | total_steps = 1 423 | for epoch in range(num_epochs): 424 | # add epoch number with corresponding step number 425 | tbwriter.add_scalar('epoch', epoch, total_steps) 426 | for i, sample_batch_pairs in enumerate(random_data_loader): 427 | # using the sample batch pair, split into 428 | # batch of combined pairs, clean signals, and noisy signals 429 | batch_pairs_var, clean_batch_var, noisy_batch_var = split_pair_to_vars(sample_batch_pairs) 430 | 431 | # latent vector - normal distribution 432 | z = sample_latent() 433 | 434 | ##### TRAIN D ##### 435 | # TRAIN D to recognize clean audio as clean 436 | # training batch pass 437 | outputs = discriminator(batch_pairs_var, ref_batch_var) # out: [n_batch x 1] 438 | clean_loss = torch.mean((outputs - 1.0) ** 2) # L2 loss - we want them all to be 1 439 | 440 | # TRAIN D to recognize generated audio as noisy 441 | generated_outputs = generator(noisy_batch_var, z) 442 | disc_in_pair = torch.cat((generated_outputs.detach(), noisy_batch_var), dim=1) 443 | outputs = discriminator(disc_in_pair, ref_batch_var) 444 | noisy_loss = torch.mean(outputs ** 2) # L2 loss - we want them all to be 0 445 | d_loss = 0.5 * (clean_loss + noisy_loss) 446 | 447 | # back-propagate and update 448 | discriminator.zero_grad() 449 | d_loss.backward() 450 | d_optimizer.step() # update parameters 451 | 452 | ##### TRAIN G ##### 453 | # TRAIN G so that D recognizes G(z) as real 454 | z = sample_latent() 455 | generated_outputs = generator(noisy_batch_var, z) 456 | gen_noise_pair = torch.cat((generated_outputs, noisy_batch_var), dim=1) 457 | outputs = discriminator(gen_noise_pair, ref_batch_var) 458 | 459 | g_loss_ = 0.5 * torch.mean((outputs - 1.0) ** 2) 460 | # L1 loss between generated output and clean sample 461 | l1_dist = torch.abs(torch.add(generated_outputs, torch.neg(clean_batch_var))) 462 | g_cond_loss = g_lambda * torch.mean(l1_dist) # conditional loss 463 | g_loss = g_loss_ + g_cond_loss 464 | 465 | # back-propagate and update 466 | generator.zero_grad() 467 | g_loss.backward() 468 | g_optimizer.step() 469 | 470 | # print message and store logs per 10 steps 471 | if (i + 1) % 20 == 0: 472 | print( 473 | 'Epoch {}\t' 474 | 'Step {}\t' 475 | 'd_loss {:.5f}\t' 476 | 'd_clean_loss {:.5f}\t' 477 | 'd_noisy_loss {:.5f}\t' 478 | 'g_loss {:.5f}\t' 479 | 'g_loss_cond {:.5f}' 480 | .format(epoch + 1, i + 1, d_loss.item(), clean_loss.item(), 481 | noisy_loss.item(), g_loss.item(), g_cond_loss.item())) 482 | 483 | ### Functions below print various information about the network. Uncomment to use. 484 | # print('Weight for latent variable z : {}'.format(z)) 485 | # print('Generated Outputs : {}'.format(generated_outputs)) 486 | # print('Encoding 8th layer weight: {}'.format(generator.module.enc8.weight)) 487 | 488 | # record scalar data for tensorboard 489 | tbwriter.add_scalar('loss/d_loss', d_loss.item(), total_steps) 490 | tbwriter.add_scalar('loss/d_clean_loss', clean_loss.item(), total_steps) 491 | tbwriter.add_scalar('loss/d_noisy_loss', noisy_loss.item(), total_steps) 492 | tbwriter.add_scalar('loss/g_loss', g_loss.item(), total_steps) 493 | tbwriter.add_scalar('loss/g_conditional_loss', g_cond_loss.item(), total_steps) 494 | 495 | # save sampled audio at the beginning of each epoch 496 | if i == 0: 497 | z = sample_latent() 498 | fake_speech = generator(fixed_test_noise, z) 499 | fake_speech_data = fake_speech.data.cpu().numpy() # convert to numpy array 500 | fake_speech_data = emph.de_emphasis(fake_speech_data, emph_coeff=0.95) 501 | 502 | for idx in range(num_gen_examples): 503 | generated_sample = fake_speech_data[idx] 504 | gen_fname = test_noise_filenames[idx] 505 | filepath = os.path.join( 506 | gen_data_path, '{}_e{}.wav'.format(gen_fname, epoch)) 507 | # write to file 508 | wavfile.write(filepath, sample_rate, generated_sample.T) 509 | # show on tensorboard log 510 | tbwriter.add_audio( 511 | '{}/{}'.format(epoch, gen_fname), 512 | generated_sample.T, 513 | total_steps, 514 | sample_rate) 515 | 516 | total_steps += 1 517 | 518 | # save various states 519 | state_path = os.path.join(checkpoint_path, 'state-{}.pkl'.format(epoch + 1)) 520 | state = { 521 | 'discriminator': discriminator.state_dict(), 522 | 'generator': generator.state_dict(), 523 | 'g_optimizer': g_optimizer.state_dict(), 524 | 'd_optimizer': d_optimizer.state_dict(), 525 | } 526 | torch.save(state, state_path) 527 | 528 | ### Can be loaded using, for example: 529 | # states = torch.load(state_path) 530 | # discriminator.load_state_dict(state['discriminator']) 531 | 532 | tbwriter.close() 533 | print('Finished Training!') 534 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | audioread==2.1.6 2 | decorator==4.3.2 3 | joblib==0.13.2 4 | librosa==0.6.3 5 | llvmlite==0.27.1 6 | numba==0.42.1 7 | numpy==1.16.1 8 | protobuf==3.6.1 9 | PyYAML==4.2b1 10 | resampy==0.2.1 11 | scikit-learn==0.20.2 12 | scipy==1.2.1 13 | six==1.12.0 14 | tensorboardX==1.6 15 | torch==1.0.1.post2 16 | -------------------------------------------------------------------------------- /vbnorm.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.autograd import Variable 4 | from torch.nn.parameter import Parameter 5 | from torch.nn.modules import Module 6 | 7 | 8 | class VirtualBatchNorm1d(Module): 9 | """ 10 | Module for Virtual Batch Normalization. 11 | 12 | Implementation borrowed and modified from Rafael_Valle's code + help of SimonW from this discussion thread: 13 | https://discuss.pytorch.org/t/parameter-grad-of-conv-weight-is-none-after-virtual-batch-normalization/9036 14 | """ 15 | def __init__(self, num_features: int, eps: float=1e-5): 16 | super().__init__() 17 | # batch statistics 18 | self.num_features = num_features 19 | self.eps = eps # epsilon 20 | self.ref_mean = self.register_parameter('ref_mean', None) 21 | self.ref_mean_sq = self.register_parameter('ref_mean_sq', None) 22 | 23 | # define gamma and beta parameters 24 | gamma = torch.normal(mean=torch.ones(1, num_features, 1), std=0.02) 25 | self.gamma = Parameter(gamma.float().cuda(async=True)) 26 | self.beta = Parameter(torch.cuda.FloatTensor(1, num_features, 1).fill_(0)) 27 | 28 | def get_stats(self, x): 29 | """ 30 | Calculates mean and mean square for given batch x. 31 | Args: 32 | x: tensor containing batch of activations 33 | Returns: 34 | mean: mean tensor over features 35 | mean_sq: squared mean tensor over features 36 | """ 37 | mean = x.mean(2, keepdim=True).mean(0, keepdim=True) 38 | mean_sq = (x ** 2).mean(2, keepdim=True).mean(0, keepdim=True) 39 | return mean, mean_sq 40 | 41 | def forward(self, x, ref_mean: None, ref_mean_sq: None): 42 | """ 43 | Forward pass of virtual batch normalization. 44 | Virtual batch normalization require two forward passes 45 | for reference batch and train batch, respectively. 46 | The input parameter is_reference should indicate whether it is a forward pass 47 | for reference batch or not. 48 | 49 | Args: 50 | x: input tensor 51 | is_reference(bool): True if forwarding for reference batch 52 | Result: 53 | x: normalized batch tensor 54 | """ 55 | mean, mean_sq = self.get_stats(x) 56 | if ref_mean is None or ref_mean_sq is None: 57 | # reference mode - works just like batch norm 58 | mean = mean.clone().detach() 59 | mean_sq = mean_sq.clone().detach() 60 | out = self._normalize(x, mean, mean_sq) 61 | else: 62 | # calculate new mean and mean_sq 63 | batch_size = x.size(0) 64 | new_coeff = 1. / (batch_size + 1.) 65 | old_coeff = 1. - new_coeff 66 | mean = new_coeff * mean + old_coeff * ref_mean 67 | mean_sq = new_coeff * mean_sq + old_coeff * ref_mean_sq 68 | out = self._normalize(x, mean, mean_sq) 69 | return out, mean, mean_sq 70 | 71 | def _normalize(self, x, mean, mean_sq): 72 | """ 73 | Normalize tensor x given the statistics. 74 | 75 | Args: 76 | x: input tensor 77 | mean: mean over features. it has size [1:num_features:] 78 | mean_sq: squared means over features. 79 | 80 | Result: 81 | x: normalized batch tensor 82 | """ 83 | assert mean_sq is not None 84 | assert mean is not None 85 | assert len(x.size()) == 3 # specific for 1d VBN 86 | if mean.size(1) != self.num_features: 87 | raise Exception( 88 | 'Mean size not equal to number of featuers : given {}, expected {}' 89 | .format(mean.size(1), self.num_features)) 90 | if mean_sq.size(1) != self.num_features: 91 | raise Exception( 92 | 'Squared mean tensor size not equal to number of features : given {}, expected {}' 93 | .format(mean_sq.size(1), self.num_features)) 94 | 95 | std = torch.sqrt(self.eps + mean_sq - mean**2) 96 | x = x - mean 97 | x = x / std 98 | x = x * self.gamma 99 | x = x + self.beta 100 | return x 101 | 102 | def __repr__(self): 103 | return ('{name}(num_features={num_features}, eps={eps}' 104 | .format(name=self.__class__.__name__, **self.__dict__)) 105 | --------------------------------------------------------------------------------